分散型光ファイバー侵入検知のための動的プロトタイプによる説明可能な少数ショット学習

Reliable intrusion detection is critical for modern infrastructure security, yet it faces two fundamental challenges: scarcity of labeled samples and lack of model interpretability. Distributed optical fiber vibration sensing (DVS) systems are promising for perimeter security but perform poorly when only a few intrusion samples are available. Most deep models also lack transparency and trustworthiness. To address these issues, we propose an explainable dual-branch feature fusion dynamic class center prototypical network (DBFF-DC-ProtoNet). The framework employs a lightweight dual-branch 1-D ResNet to extract complementary temporal and time-frequency representations from raw signals and discrete wavelet transform (DWT) features, which are fused to form more discriminative class prototypes. A dynamic class center update strategy with a novel loss function is further introduced to enhance intra-class compactness and inter-class separability in few-shot conditions. In addition, an explainability module integrates prototype-based class activation mapping (Proto-CAM) and case-based reasoning, offering both fine-grained attribution of key signal segments and intuitive retrieval of similar historical cases. Extensive experiments on a self-collected dataset and a public benchmark confirm the effectiveness of our approach, achieving 97.22% and 98.33% accuracy under the 5-shot setting. These results demonstrate that DBFF-DC-ProtoNet effectively bridges few-shot learning with interpretability, providing a practical and trustworthy solution for DVS-based intrusion detection.